Implement materialize_mixed_upgrade: per-year ResStock data for HP adoption

Assigns baseline buildings to upgrades using a monotonic random-seed
allocation (contiguous shuffled-rank bands, one per upgrade), then writes
per-year materialized data that run_scenario.py can consume unchanged:

- metadata-sb.parquet: rows from each upgrade's metadata filtered to
  the buildings assigned to it for that year, combined into one file.
- loads/: symlinks pointing each building to the correct upgrade's
  load parquet ({bldg_id}-{N}.parquet), discovered from the source
  directory rather than hard-coding a filename format.

Buildings with postprocess_group.has_hp == True in the upgrade-0
baseline are excluded from random assignment and pinned to upgrade 0
in every year (they already have a heat pump and should not be
re-assigned to a new HP upgrade).

Also writes scenario_assignments.csv for reference (bldg_id +
one column per run year showing the assigned upgrade).

Validates the adoption config with buildstock_fetch.scenarios.validate_scenario
before any I/O; errors clearly if required upgrade directories are missing.

Author: sherryzuo